MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering

   

MLExpert is a 35B Mixture-of-Experts agentic model for autonomous machine learning engineering. It organizes experiments, writes and revises code, and improves solutions through sustained interaction with execution environments. Hierarchical agentic supervised fine-tuning (SFT) develops general tool-use and domain-specific optimization skills, while Windowed Interaction Reinforcement Learning (WIRL) trains multi-turn decisions using feedback from completed experiments.

Highlights

• Long-horizon optimization: Plans experiments, interprets execution feedback, revisits previous solutions, and selects submission candidates within a persistent workspace.
• MLE performance: Achieves a 72.7% medal rate on MLE-Bench Lite, matching GLM-5.3 and Kimi-K3 under the MLExpert Harness.
• General-purpose harness compatibility: Achieves a 69.7% medal rate with Claude Code, compared with 42.4% for its base model.
• Scientific task optimization: Matches published scientific results on seven of ten NatureBench Lite tasks and surpasses them on five.

Method

Model-directed MLE

The model selects subsequent actions based on experimental states and execution feedback. The interaction harness maintains candidate solutions and experiment history throughout long-horizon optimization.

Task construction and post-training

The Kaggle-MLE Factory constructs executable machine learning tasks. The interaction harness supports model-directed experimentation and teacher trajectory collection. Hierarchical agentic SFT transfers general coding and tool-use capabilities to MLE through outcome-guided supervision. WIRL groups decisions between successive experiment results into execution windows and learns from intermediate execution feedback through asynchronous collection and policy correction.

Performance

MLE-Bench Lite

MLE-Bench Lite comprises 22 historical Kaggle competitions for end-to-end machine learning engineering. The main metric is the percentage of tasks receiving any medal. Under the MLExpert Harness, MLExpert raises its base model's medal rate from 37.9% to 72.7%. With Claude Code, the rate improves from 42.4% to 69.7%.

NatureBench Lite

NatureBench Lite comprises ten scientific experimentation tasks. M denotes Match-SOTA, and S denotes Surpass-SOTA. With Claude Code, MLExpert raises the Surpass-SOTA rate from 0.0% to 50.0% and the Match-SOTA rate from 40.0% to 70.0% relative to its base model.

Kaggle-MLE Factory

The Kaggle-MLE Factory builds executable MLE tasks from user-supplied Kaggle competition slugs, including downloading, data analysis, LLM-generated preparation and scoring scripts, validation, and repair. Kaggle credentials and the model service are user-configurable. Generated data splits and scoring code require human review, especially for complex competition metrics. Start with its offline example, which runs without Kaggle credentials or a model service.

Usage

MLExpert model weights are available on Hugging Face. The commands below serve an OpenAI-compatible API at http://localhost:8000/v1 with tensor parallelism across two GPUs.

vLLM

uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate
uv pip install vllm --torch-backend=auto

Text generation

vllm serve InternScience/MLExpert \
    --host 127.0.0.1 \
    --port 8000 \
    --tensor-parallel-size 2 \
    --max-model-len 262144 \
    --reasoning-parser qwen3 \
    --language-model-only

Tool calling

vllm serve InternScience/MLExpert \
    --host 127.0.0.1 \
    --port 8000 \
    --tensor-parallel-size 2 \
    --max-model-len 262144 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --language-model-only

SGLang

uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate
uv pip install sglang

Text generation

python -m sglang.launch_server \
    --model-path InternScience/MLExpert \
    --host 127.0.0.1 \
    --port 8000 \
    --tp-size 2 \
    --context-length 262144 \
    --reasoning-parser qwen3 \
    --language-model-only

Tool use

python -m sglang.launch_server \
    --model-path InternScience/MLExpert \
    --host 127.0.0.1 \
    --port 8000 \
    --tp-size 2 \
    --context-length 262144 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder \
    --language-model-only

Sampling parameters

Parameter	Value
temperature	1.0
top_p	0.95
top_k	20
min_p	0.0
presence_penalty	1.5
repetition_penalty	1.0

These parameters apply to both thinking and non-thinking modes. Set chat_template_kwargs.enable_thinking to true to enable thinking or false to disable it. Machine-readable profiles are provided in the model repository's inference_profiles.json.